With a fix for bug 5309, such as the one discussed at https://gerrit.wikimedia.org/r/121255/, it’s entirely possible that a user might get a CAPTCHA with illegible diacritics. Diacritics in Latin alphabets can look identical to one another when distorted, for example i í ì ỉ, or ó ơ.
For better usability, ConfirmEdit should display a CAPTCHA containing diacritics but require the user to enter the characters without diacritics. There’s a third-party module called Unidecode that does a decent job of accent folding.
One tradeoff would be that such CAPTCHAs might be easier for a bot to crack. There’s also the issue that a character like Ê might be considered a base letter in one language (as in Vietnamese) but a letter with a diacritic in another (Portuguese).